How to Choose a Pharmacogenetic Testing Company (and When to Skip the Panel)

There is no single best pharmacogenetic testing company, and the question most buyers ask (“Tempus or GeneSight?”) is the wrong one. What matters is which star alleles a lab genotypes, whether it resolves copy number and hybrid structures at CYP2D6, whether it returns your raw data, and whether its interpretation follows Clinical Pharmacogenetics Implementation Consortium (CPIC) guidelines or a proprietary scoring algorithm you cannot audit. Our recommendation for a technically fluent reader: get 30x whole-genome sequencing, call diplotypes yourself with PharmCAT plus a dedicated CYP2D6 caller, and use a CLIA-certified panel as an orthogonal confirmation of any result you or a clinician would act on. Any decision about a medication belongs to a prescribing clinician or a pharmacist, and ideally one with pharmacogenomics training.

What a pharmacogenetic test measures

A PGx test does not sequence “your drug response.” It genotypes a defined set of positions, assembles them into haplotypes named with star () nomenclature curated by PharmVar (CYP2C192, CYP2D6*4, and so on), pairs the two haplotypes into a diplotype, and then translates that diplotype into a metabolizer phenotype using a lookup table. CYP2D6 adds an activity score layer: each allele carries a value (0 for *4, 0.25 or 0.5 for *10 and *41 depending on the version, 1.0 for *1), the two are summed, and the sum maps to poor, intermediate, normal, or ultrarapid metabolizer.

Three things follow. First, a test can only find variants it interrogates, so panel content is the single most important specification a vendor publishes. Second, translation tables change: CPIC revised the CYP2D6*10 activity value, which moved a population of people between phenotype categories without any change to their DNA. Third, the report is a derived artifact. If you hold the variant calls, you can re-derive it later against a newer table. If you hold only a PDF, you cannot.

The American College of Medical Genetics and Genomics published a technical standard specifying minimum recommended allele sets per gene for clinical PGx testing and reporting, along with requirements for how diplotypes and phenotypes are represented.¹ Use that document as the checklist when comparing vendor menus. A lab that tests fewer alleles than the ACMG minimum for CYP2C19 or CYP2D6 is not comparable to one that meets it, regardless of marketing.

CYP2D6 and DPYD are the tests of the test

CYP2D6 sits next to two pseudogenes, CYP2D7 and CYP2D8, with high sequence identity. Real people carry whole-gene deletions (*5), duplications and multiplications of functional or nonfunctional alleles, and CYP2D6-CYP2D7 hybrid structures such as *13 and *68 that arise from unequal crossover. A genotyping array or a simple SNP panel that reports “*1/*4” may be looking at a duplicated *4 or a hybrid it cannot see. This is where vendors separate: ask whether they perform copy number analysis (typically long-range PCR or a dedicated CNV assay) and which hybrid alleles they report.

DPYD is the other discriminator, and it matters because dihydropyrimidine dehydrogenase deficiency drives severe fluoropyrimidine toxicity. The four variants with the strongest evidence include c.1129-5923C>G (HapB3), which is intronic and roughly six kilobases from the nearest exon. Exome sequencing misses it unless the capture kit was designed to include it, and many SNP panels omit it. A PGx menu that lists “DPYD” without listing the variants is not informative.

Next-generation sequencing handles this class of problem better than targeted genotyping because it interrogates the whole locus rather than a fixed variant list, though it brings its own complications in mapping and phasing across the CYP2D6 region.² Long-read platforms (PacBio HiFi, Oxford Nanopore) resolve CYP2D6 structural haplotypes and HLA alleles more cleanly than short reads, and if a vendor offers long-read PGx it is a real technical advantage rather than a marketing one.

How the companies differ

Hospital and reference laboratories (ARUP, Mayo Clinic Laboratories, academic CLIA labs) publish gene and allele menus, report CPIC diplotypes and phenotypes, and generally restrict themselves to genes with strong evidence. They are the most auditable option and usually the cheapest per gene. They typically do not return raw data, and ordering requires a clinician.

Broad clinical panels sold to prescribers (OneOme RightMed, Genomind, and similar) cover twenty to thirty genes with interpretation layered over CPIC and FDA labeling. The value here is workflow integration rather than novel measurement. Ordering friction is a real cost in practice, and standardized ordering interfaces are one of the few demonstrated ways to reduce it.³

Combinatorial psychiatric panels, of which GeneSight is the best known, combine multiple gene results into a proprietary categorical recommendation. The genotyping underneath is ordinary, but the combinatorial algorithm is not a CPIC product and its inputs and weights are not fully disclosed. The evidence base for pharmacogenetic testing overall is uneven across drug-gene pairs, and clinical utility trials have produced mixed results depending on endpoint and design, which is a central reason adoption has been slower than the biology would suggest.⁴ We would treat a combinatorial score as a hypothesis, not a result, and we would ask for the underlying diplotypes.

Direct-to-consumer PGx products are convenient and frequently under-specified. Reviews of the DTC pharmacogenomic market have identified recurring problems with allele coverage, interpretive transparency, and the absence of clinical follow-up, and have proposed concrete strategies to address them.⁵ If you buy one, buy it for the raw data.

Sequencing companies such as Tempus sit in a different category: their core business is oncology sequencing, and PGx is an adjunct. Comprehensive whole-genome assays in the oncology setting have been shown to recover clinically relevant alterations concordant with targeted panels while adding findings those panels miss, which is the general argument for sequencing the locus instead of genotyping a list.⁶

Calling your own diplotypes from WGS

Start with 30x short-read WGS aligned to GRCh38, as a CRAM or BAM plus a VCF. The reference build matters: PharmCAT’s position definitions are GRCh38-based, and a GRCh37 VCF will fail or silently mismatch.

Run the bundled pipeline:

pharmcat_pipeline sample.vcf.gz -o pgx/ -bf sample -research cyp2d6

The pipeline runs three stages. The preprocessor normalizes representation (left-aligns indels, splits multi-allelics, restricts to PharmCAT positions). The named allele matcher assigns diplotypes. The phenotyper and reporter produce sample.match.json, sample.phenotype.json, and an HTML report with CPIC and DPWG recommendations.

Two flags deserve care. --absent-to-ref (-ar) tells PharmCAT to treat positions missing from your VCF as reference. That is correct for a gVCF-derived callset with verified coverage and wrong for a sparse VCF, where it will manufacture confident wild-type calls at positions you never sequenced. Check per-position depth first: samtools depth -b pharmcat_positions.bed sample.cram and confirm at least 10x to 20x at every DPYD, TPMT, and NUDT15 position before trusting a reference call.

For CYP2D6, do not rely on the VCF. Run Cyrius directly on the aligned reads, which models CYP2D6-CYP2D7 hybrids and copy number:

python star_caller.py --manifest cram.txt --genome 38 --prefix sample --outDir cyp2d6/ --threads 8

Aldy 4 and StellarPGx are reasonable alternatives, and disagreement among them is informative: concordant calls from Cyrius and Aldy are trustworthy, discordant ones should go to a CLIA lab with a dedicated CYP2D6 assay. Feed the final call back to PharmCAT as an outside call file with -po.

Expect specific failure modes. HLA-B57:01 and HLA-B15:02 need an HLA typer such as OptiType or HLA-LA rather than a star-allele caller. G6PD is X-linked, so hemizygous males need different phenotype logic. CYP2C19*17 and *2 in cis produce diplotypes that short-read data cannot always phase.

Cost and what to buy

Self-pay prices for reference-lab PGx panels commonly land in the low hundreds of dollars, broad commercial panels higher, and insurance coverage varies by drug-gene pair and indication. Clinical-grade 30x WGS with returned CRAM and VCF is now in a comparable range and gives you every PGx locus plus everything else, at the cost of doing the diplotype calling yourself.

Our view: buy sequencing, keep the CRAM, and run PharmCAT plus Cyrius. Then, if a specific result would inform a prescribing decision, confirm that gene in a CLIA laboratory and take both to a clinician or pharmacogenomics pharmacist. Research-pipeline calls are not clinical results, and the gap matters most precisely for the genes where the stakes are highest.

Questions people also ask

Is Tempus or GeneSight better? They solve different problems. Tempus is a sequencing company whose PGx output derives from broader molecular profiling, while GeneSight is a fixed psychiatric panel with a proprietary combinatorial interpretation. If you want auditable diplotypes and data you can re-analyze, sequencing-based options are preferable.

Which genetic testing is the most accurate? Accuracy is gene-specific. For DPYD and TPMT, a well-designed panel or WGS both perform well. For CYP2D6, only assays with explicit copy number and hybrid detection are accurate, and long-read sequencing is the current technical ceiling.

What is the typical cost of a pharmacogenetic test? Reference-lab panels are usually in the low hundreds of dollars self-pay. Commercial multi-gene panels cost more, and coverage depends on the drug and indication.

What is the best pharmacogenomic test? The one whose allele menu meets the ACMG minimum sets, resolves CYP2D6 structural variation, reports CPIC diplotypes and phenotypes rather than proprietary categories, and returns your raw data.¹

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. Marwan K. Tayeh, Andrea Gaedigk, Matthew P. Goetz, et al. Clinical pharmacogenomic testing and reporting: A technical standard of the American College of Medical Genetics and Genomics (ACMG). Genetics in Medicine, 2022. https://doi.org/10.1016/j.gim.2021.12.009 ↩ ↩²

2. Alireza Tafazoli, Henk-Jan Guchelaar, Wojciech Miltyk, et al. Applying Next-Generation Sequencing Platforms for Pharmacogenomic Testing in Clinical Practice. Frontiers in Pharmacology, 2021. https://doi.org/10.3389/fphar.2021.693453 ↩

3. Kashmira Rajendra Sawant, Howard Anton, Kinson Ho, et al. Real-world evaluation of the reduction in complexity of ordering DNA sequencing tests through the adoption of Fast Healthcare Interoperability Resources (FHIR) standards. Journal of Clinical Oncology, 2025. https://doi.org/10.1200/jco.2025.43.16_suppl.e13719 ↩

4. Nihal El Rouby, Julie A. Johnson. Pharmacogenetic Testing — Evidence, Challenges, and Pathways to Adoption. NEJM Evidence, 2025. https://doi.org/10.1056/evidra2400343 ↩

5. Alireza Tafazoli, Rama Krishna Guggilla, Zahra Kamel-Koleti, et al. Strategies to Improve the Clinical Outcomes for Direct-to-Consumer Pharmacogenomic Tests. Genes, 2021. https://doi.org/10.3390/genes12030361 ↩

6. Robert Huether, Derick Hoskinson, Pavana Anur, et al. A Comprehensive Whole Genome Sequencing Assay Provides Robust Characterization of Clinically Relevant Genomic Alterations across Myeloid Malignancies Concordant with Matched Results from Targeted DNA, Whole Transcriptome RNA and Cytogenetic Profiling. Blood, 2024. https://doi.org/10.1182/blood-2024-199194 ↩